Week 09 · lesson

Lab: Compare Attention Allocations

Mission

Run the local Attention Allocation Lab. Apply two queries to the same three keys, reproduce softmax normalization, and export an Attention Allocation Record.

Guided start

Select Object query [1,0] and run attention.
Confirm that robot has the largest weight.
Verify that all three weights sum to 1.000 after displayed rounding.
Switch only the query to Action query [0,1].
Explain why safely now receives the largest weight.

Procedure

Preserve the fixed keys and both query vectors.
Predict the top-weighted token for each query.
Run both calculations and preserve every dot product and exponential term.
Verify normalization from the displayed denominator.
Compare the allocations without inventing meanings for vector dimensions.
Preserve your first explanation claim and write a bounded correction.
Complete Observation, Measurement, Decision, and Limitation.
Export the record and confirm it names ai-for-beginners-attention-allocation-lab.

Success criteria and boundary

both queries and all keys remain visible;
each weight can be reproduced from its score and denominator;
the displayed weights sum to 1 within rounding;
the comparison changes only the query; and
the limitation rejects probability, causal explanation, training, safety, and whole-model-performance claims.

Use only supplied vectors. The Lab contacts no transformer or external service.

Robotnix Lab · contract v1

Attention Allocation Lab

Calculate single-head attention weights under two queries and bound claims about transformer explanations.

Help and boundaries

A deterministic browser simulation uses invented two-dimensional vectors and contacts no model or service.

No personal data, private prompts, or API keys. This Lab makes no hidden remote writes and does not persist your work after the page session.

No result generates external text, inspects a live model, or authorizes action.

Lab draft is active for this page session only.

Objectives

Calculate query-key dot products and softmax weights.
Compare allocations while changing only the query.
Separate attention allocation from probability, causal explanation, reasoning, and safety.

Procedure

Inspect vectorsRecord both queries and the three fixed key vectors.
Calculate weightsPreserve dot products, exponential terms, denominator, and normalized weights.
Compare queriesRun both queries and identify how the allocation changes.
Bound attention claimSeparate supplied arithmetic from probability, causal explanation, reasoning, safety, and trained-model behavior.

Workbench

Calculate the allocation before interpreting attention.

Query vector

Select a supplied query and run the deterministic attention calculation.

Boundary: These invented vectors simulate one simplified head. A weight is not a calibrated probability, causal explanation, proof of reasoning, safety judgment, or evidence about a trained transformer.

Text and static fallback

Every query, key, score, exponential term, denominator, weight, sum, and boundary appears in accessible text and tables.

Object query: robot about 0.425; action query: safely about 0.514; each unrounded allocation sums to one.

Evidence record

Record only the evidence requested. Do not enter names, personal information, private prompts, or credentials. This draft stays in memory only unless you explicitly export it.

Vector declaration *Record both query vectors, all key vectors, and the simplified-head boundary.Attention calculations *Record every score, exponential term, denominator, and weight for both queries.Attention claim repair *Preserve and repair one attention-explains-the-model claim.Limitation *State what invented single-head weights cannot establish about a trained transformer.

Success criteria

Both queries, all keys, and the teaching boundary are explicit.
Every weight is reproducible from displayed operands.
The correction follows supplied allocation evidence only.
The limitation rejects unsupported probability, explanation, reasoning, safety, and performance claims.

Complete every procedure step and required evidence field before export.

Import evidence JSON

Vocabulary lab

Flip the idea, not just the card

Explain the term before you reveal the back. Then compare your explanation with the definition, example, and warning.

Read all terms without animation

dot-product score: The sum of matching query-key component products. Example: Object dot robot equals 0.9. Do not confuse it with: It is unnormalized before softmax.
attention weight: A normalized allocation assigned to one position for one query and head. Example: Robot receives about 0.425 under the object query. Do not confuse it with: It is not a complete causal explanation.
attention head: One set of query, key, and value projections producing an attention result. Example: The Lab simulates one simplified head. Do not confuse it with: Real transformers commonly use multiple heads.
normalization denominator: The sum of all exponentiated scores used by softmax. Example: The object-query denominator is about 5.790. Do not confuse it with: It depends on every candidate position.
Attention Allocation Record: Evidence preserving vectors, calculations, comparison, claim repair, and limitation. Example: The Lab export. Do not confuse it with: It is not evidence from a trained transformer.